Program Information
Using Patient-Specific Factors to Predict Intra-Fraction Motion in Prostate Cancer Patients with Machine-Learning
A Coathup*, P Basran , BC Cancer Agency, Victoria, BC
Presentations
SU-I-GPD-J-56 (Sunday, July 30, 2017) 3:00 PM - 6:00 PM Room: Exhibit Hall
Purpose: To develop a machine-learning system which predicts the extent of intra-fraction prostate motion for prostate cancer patients using patient-specific factors, such as body mass-index, PSA, and other demographic data.
Methods: N=21 prostate cancer patients with intra-fraction motion data from gold fiducial markers were used for machine-learning training and validation. Leave One Out Cross Validation Method (LOO-CV) was used to train the model and evaluate its performance in predicting a patient’s intra-fraction prostate motion. Predicted motion was compared with observed motion. Additional simulations were performed using a model to evaluate the projected performance of the algorithm.
Results: The machine-learning system predicted both the mean and the standard deviation of a prostate cancer patient’s intra-fraction motion to within a maximum of 0.8mm mean absolute error along all of the vertical, longitudinal, and lateral directions. Simulation results suggest a dramatically increased accuracy with increased number of patients especially up to N > 50.
Conclusion: We have developed a strategy and algorithm for predicting the extent of tumor motion using machine-learning methods with patient-specific factors as inputs. The LOO-CV approach is an appropriate measure of model performance and simulates a clinically implementable approach. Increased intra-fraction motion prediction accuracy is expected as N increases. This approach can be easily adapted to other tumor sites permitting the opportunity for patient-specific PTV margins.
Funding Support, Disclosures, and Conflict of Interest: This work is partially funded by the Vancouver Island Prostate Centre-West Coast Ride To Live Grant
Contact Email: